dimensional-multiplexingActiveEvidence cut / 23 Aug 2026

Dimensional Multiplexing

Can reconstruction compute replace persistent model memory while improving capability per deployed byte?

Current answer

The tested construction does not improve language-model capability per deployed byte over conventional controls.

Current direction

Novel mechanisms and memory objects with new, discriminating predictions.

~4.1×storage of matched dense modelBounded synthetic composition result
≤0.996×storage multiplier bound at 50MOutside the measured dense frontier
1.74×whole-model structural compressionPruning remains marginally better
2.5×recovery-token savingMechanism-independent process result
01 / Overview

Store less. Reconstruct more.

Dimensional Multiplexing started from a systems constraint rather than a compression target. Modern models can become memory-bound long before the available arithmetic is exhausted. If useful weight states share enough structure, some of that persistent state may be replaceable with reconstruction compute.

The project established that the construction is mechanically real. Address-selected transforms create distinct logical weights, the byte accounting is exact, and the representation can compose structurally with quantization. On one balanced synthetic composition task it matched a dense INT4 model using about four times its storage.

The general language-model claim did not survive. Across controlled transformer tests, pretrained compression and mechanism probes, the derived logical width behaves primarily like the number of stored degrees of freedom rather than a larger independently learned model.

Current result

The structural compression is real. The general language capability advantage is not.

02 / Mechanism

A smaller stored base, expanded at use.

The core construction derives multiple logical weight planes from shared stored state:

W[q, j]=Tq(Bj)+R[q, j]
Bj

Shared stored base state.

Tq

Cheap address-selected transform used to derive a plane.

R[q,j]

Optional plane-specific residual state restoring local individuality.

The controlled transformer changes only the two-matrix GELU FFN. Attention, embeddings, RoPE, RMSNorm, residual paths and the language head remain conventional. The tested line uses eight all-active planes, two shared bases, learned Givens butterflies, optional two-sided low-rank residuals and INT4 QAT.

Deployment contract.Expanded planes cannot become persistent weights. Reconstruction must occur inside the consuming operation. Materialising the expanded tensor removes the storage advantage.
03 / Key evidence

The language result is negative under the comparison that matters.

The primary functional comparison is against a measured dense INT4 frontier at equal deployed bytes. Matched training tokens and matched tokens per stored parameter remain separate comparison classes.

20M controlled transformer

ArmDeployed bytesTest loss
Dense INT4 / FFN 102412.475 MB3.33446
Dense INT4 / FFN 216 / equal-byte control7.309 MB3.59289
Dimensional INT4 / 10% residual / uniform7.347 MB3.62284

The dimensional arm was 0.02995 nats worse despite being 0.5% larger and doing 54% more arithmetic. At matched training tokens the same comparison was +0.01752 nats. Both comparison classes agree: logical width did not buy more capability than storing a narrower dense FFN.

At 50M, the tight equal-byte pair remained unresolved and the storage multiplier was bounded at no more than 0.996x because the result fell outside the measured dense frontier. There is no advantage to trend. Across the eight-cell residual-by-mixer sweep, more individuality improved absolute loss while capability per byte fell from 1.000x to 0.942x.

Pretrained conversion

The representation was also tested as a way to compress an existing model. Under the same fitting and recovery treatment, the frontier for SmolLM2-135M was:

RouteCompressionFinal bits / UTF-8 byte
INT41.000x0.8961
INT31.151x0.9350
INT21.357x1.0497
Structured pruning1.742x1.0894
Dimensional sharing1.743x1.0951

Dimensional sharing reaches beyond the practical low-bit quantization frontier, but it does not own that advantage. Structured pruning reaches essentially the same ratio and is marginally better and simpler.

Treatment parity changed the conclusion.Earlier apparent advantages over pruning and INT2 were withdrawn after those controls received the same fitting and recovery process.
04 / What survived

The project produced more than one kind of result.

Structural

Exact byte accounting

Stored values, scales, transforms, residuals, metadata, biases and padding are charged. Sharing and quantization compose structurally.

Functional

Structured composition can benefit

The strongest bounded small-task result matched a dense INT4 model using about 4.1x its deployed storage. It did not transfer to language.

Mechanistic

Local freedom matters

Selective individuality and learned coordinate mixing can help when the task contains compatible structure. The useful rate is task-dependent.

Methodological

Recovery became much cheaper

Sequential activation calibration, teacher distillation and a cosine schedule reached a better loss after 30M recovery tokens than the previous process reached after 75M.

Reusable engineering results

  • Weight-space reconstruction is preferable to expanding activations for this implementation.
  • Whole-step integration gates matter more than isolated kernel microbenchmarks.
  • CUDA Graph capture, fused AdamW and selected Triton epilogues materially improved the research stack.
  • Test and payload gates caught defects that changed scientific conclusions, not merely code quality.
05 / Mechanism limits

The failure mode is now measured.

01

Learned FFN neurons are close to incoherent.

Within-layer mean absolute cosine similarity is 0.0397 against a random-direction floor of 0.0333. Across layers it is 0.0348. There is some structure, but not much.

02

Cheap transforms contain too little information.

A transform only creates a storage advantage if its description is cheaper than directly storing the weights. Richer transforms fit more structure, but their descriptions consume the byte saving. A full rotation costs 1.50x the values of one plane and still leaves 58% relative error.

03

Encoding the difference does not change the rate problem.

Across seven residual codecs and a fourfold rate range, plane-difference coding produced zero wins in 40 functioning equal-bit-budget comparisons against directly quantizing the wide weights.

Mechanistic interpretation

Cheap transforms compress because they contain little information. Because they contain little information, they cannot generate the missing independent capability. Rich transforms can describe more structure, but their descriptions consume the storage saving.

06 / Claim state

Keep the failed questions visible.

The point of retaining falsified and withdrawn avenues is not to make the project look unsuccessful. It is to stop the next iteration from quietly becoming another version of something already tested.

ClaimQuestionState
C-0001Address-selected transforms produce distinct functional weights.Supported
C-0002Address bits create exponential independent capacity.Falsified
C-0003Greater effective rank implies greater capability.Falsified
C-0004Logical expansion improves capability per deployed byte over a storage-matched dense control.Falsified
C-0005Small local individuality can help shared planes.Supported / task-bounded
C-0005Coordinate mixing is useful.Supported / task-bounded
C-0007Conditional plane mixing generalises.Falsified as a general claim
C-0007Sparse MoE routing unlocks the planes.Falsified as a general route
C-0008Dimensional Multiplexing beats dense INT4 at equal bytes on language.Falsified
C-0011The advantage improves with scale.Falsified / mis-posed
C-0006Sharing stacks beneficially with lower-bit quantization.Falsified functionally
C-0012Sharing beats pruning when converting a trained model.Falsified
C-0014Sharing along depth escapes the neuron-axis limit.Falsified
C-0015Plane differences can be encoded more cheaply than the weights.Falsified as a rate claim
C-0013Better recovery materially improves compressed models.Supported
C-0016Deployment throughput is proven.Open
C-0017A novel dimensional mechanism can improve capability per deployed byte outside the tested construction family.Open
Full negative / withdrawal register31 records
Avenue or claimDispositionWhat killed or limited it
Exponential capacity from address bitsFalsifiedAdditive span is at most b+1; derived states are correlated.
Effective rank as the capability objectiveFalsifiedHigher-rank compositional model lost to dense.
More transform stages as the primary fixFalsifiedFit improved; trained capability did not.
More shared basesFalsified as plateau fixDid not remove the storage-matched ceiling.
Instantaneous gradient conflict as an allocation ruleWeakenedConflict exists but did not predict a useful allocation.
Longer training as the universal explanationFalsifiedDid not remove the small-task plateau; later recovery gains came from process, not duration alone.
Dense-seeded continuationFalsifiedTraining-matched identity control explained the gain.
Fixed sharing rate everywhereFalsifiedResidual individuality has a task-dependent interior optimum.
Per-weight DFC as deployment lineRejectedMetadata overhead is disproportionate beside INT4.
Sparse MoE as general unlockFalsifiedTask-conditional benefit; spectral regression failure; all-active arm stronger.
Conditional mixer as universal componentFalsifiedComposition win did not transfer; language at zero residual worsened significantly.
Mixer compressionMostly failedRemoved or degraded the useful conditional behavior for little gain.
Checkpointed 63-cell transformer sweepInvalidLate-binding closure corrupted gradients.
Butterfly fit error as capability predictorFalsifiedClean out-of-sample prediction failed; ordering nearly reversed.
Block-diagonal rotationRejectedBetter fit, faster and smaller, but worse trained loss.
Dimensional INT2RejectedPays sharing and low-bit penalties plus a negative interaction.
Residual individuality restores capability per byte on languageFalsifiedAbsolute loss improves while storage multiplier falls monotonically.
Conditional mixing closes the transformer gapFalsifiedNo architecture cell crosses 1.0x.
Scaling will reveal a latent advantageFalsified to 50MFlat trend; at most 0.996x at 50M as a frontier bound. Nothing above 50M is measured.
Prune + share stackingFalsifiedSame-byte arrangements remain inside 0.0224 bits/byte.
Sharing recovers faster than pruningWithdrawnPruning had not received activation-aware fitting.
Sharing is 2.3x less damaging than INT2WithdrawnINT2 had not received the recovery process.
Per-layer rank tiltFalsified as material leverLarge zero-shot screen became +0.0018 after recovery.
Depth multiplexingFalsified0.9838x and 41% more arithmetic under exceptionally tight controls.
Plane collapse explains parityFalsifiedA 1.8x diversity change produced effectively identical loss.
Weight decay explains parityFalsifiedRemoving the asymmetric decay still produced parity.
Delta is cheaper than WFalsified as a rate claimSeven codecs; zero functioning equal-bit-budget wins.
Cluster similar neurons to ease sharingFalsifiedLower reconstruction error, worse functional loss.
Orthogonal spread is a free winWithdrawnPartial-run positional result did not replicate across grouping conditions.
PDACA recurrence/operator-diversity evidenceQuarantinedUnverified, post-hoc precursor evidence; not support for the mechanism.
Capability equals distinct learned parameter countWithdrawnWhere bytes are placed changes their value; individuality can improve loss while worsening loss per byte.
07 / Research path

The project progressively removed its own shortcuts.

01

Inception

Show that addressed transforms can create distinct functional states, then bound the capacity claim.

02

Local freedom

Add residual individuality and coordinate transforms; establish that local freedom can help without solving the general ceiling.

03

Capability per byte

Reset evaluation around whole-model deployed bytes and a measured dense frontier.

04

Transformer instrument

Build a controlled FFN-only transformer stack, exact accounting and qualified performance path.

05

Language result

Run equal-byte controls at 20M and 50M. Logical expansion does not beat stored dense capacity.

06

Pretrained conversion

Compare sharing, pruning and quantization under treatment-matched recovery.

07

Mechanism limits

Measure incoherence, transform cost, plane diversity and residual coding to explain the ceiling.

08

New mechanism search

Move away from local variations and test constructions with a genuinely different source of advantage.

08 / Current work

The next mechanism has to change the question.

The negative results define requirements for the next constructions; they do not define the end of the project.

Preregistered

Plane-count sweep

Test 4, 8 and 16 planes while holding the stored base and deployed-byte budget constant. The current prediction is a flat capability line with increasing compute. Eight planes is one measured point, not a theorem about every expansion ratio.

Open lead

Function-optimised neuron assignment

Treat plane construction as an assignment problem rather than assuming contiguous neuron order is neutral. Compare positional, random, clustered, diversity-spread and function-optimised groupings under identical recovery.

New mechanism

Correlated function families

Apply multiplexing where correlation exists before compression rather than forcing arbitrary FFN neurons into a shared representation. Candidate targets include related adapters, task experts, multilingual adaptations and temporal checkpoints.

New mechanism

Adaptive reconstruction

Move beyond a fixed reconstructed tensor. Activation-conditioned generation, recurrent refinement or conditional low-rank updates make a different prediction because additional compute produces an input-dependent function.

New object

Different memory objects

Weights may be the wrong object. KV state and intermediate activations can be more locally correlated and often dominate long-context deployment memory.

Measurement

Capability beyond mean NLL

Construct tasks with explicit sets of valid hypotheses and measure valid-hypothesis mass, diversity inside the valid set and collapse after evidence arrives. This is a separate measurement programme, not evidence for the architecture.

Deployment gate

Fused reconstruct-and-consume inference

A projected 70B INT4 decode model remains bandwidth-bound after reconstruction and suggests a possible throughput gain. This is a model, not a benchmark. No fused kernel or 70B dimensional checkpoint currently proves the deployment claim.

09 / Limits

The falsification is broad, not universal.

  • From-scratch language training was measured at 20M and 50M only.
  • SmolLM2-135M was used for pretrained compression, not an equivalent from-scratch campaign.
  • Capability is held-out FineWeb-Edu pretraining loss, not downstream utility.
  • The controlled transformer changes the FFN rather than every memory-bearing component.
  • The positive small-task composition result remains real even though it did not generalise.
  • The plane-count axis and recovered neuron-permutation lead remain open.
  • Adaptive reconstruction and different memory objects lie outside the present falsification.
  • Large-model storage and decode gains remain structural or projected, not measured deployment results.
Appropriate conclusion

Shared bases, cheap address-selected transforms and small charged corrections have not increased language-model capability per deployed byte over conventional alternatives.

The evidence does not justify the broader claim that learned computation can never trade for stored weights.
10 / Public evidence

The public record ends at the public repository boundary.

Experiment records, machine-readable results, claim dispositions and supporting artifacts are maintained in the public project repository. Experiment and claim identifiers used here correspond to immutable public records.

  • E-0011 — Transformer mechanism sweep
  • E-0012 — Transformer scale probe
  • E-0013 — Pretrained compression and recovery comparison
  • E-0015 — Depth multiplexing
  • E-0016 — Activation-aware recovery process
  • E-0018 — Plane coherence and angle decay
  • E-0019 — Plane rate-distortion

Evidence that has not crossed the public repository boundary is described as unpublished and is not referenced by an internal filename.